Three kinds of AI people often mix together.

Design focus

The useful question is not whether a system "uses AI." The useful question is what kind of work the AI is being asked to do: estimate what is likely, create something new, or help a person interact with a system in natural language.

AI is not one thing

A lot of conversations about AI get muddy because the same word is used for several different capabilities. A forecasting model, an image generator, and a chat interface can all be called AI, but they have different jobs, different risks, and different ways they should be evaluated.

That distinction matters in real systems. If the goal is to estimate demand, risk, failure, or timing, the system needs predictive strength and measurable performance. If the goal is to draft, summarize, translate, or generate examples, the system needs generative capability and review boundaries. If the goal is to let people ask questions, give instructions, or navigate a workflow, the system needs a conversational interface that stays tied to the right tools and data.

Most useful AI systems are not built from one of these categories alone. They combine them carefully.

Operating principle

Name the type of AI before deciding how much authority the system should have.

Predictive AI

Estimates what is likely to happen, what category something belongs in, or what risk a situation carries.

Generative AI

Creates new text, images, audio, code, summaries, plans, translations, or examples from patterns it has learned.

Conversational AI

Lets people interact with systems through questions, instructions, dialogue, and natural-language workflows.

Predictive AI: what is likely?

Predictive AI is the part of AI most people were using long before the current wave of chat tools. It looks at data and estimates a likely outcome: demand next month, the probability of churn, the chance of equipment failure, the best route, the next likely word, the risk level of a transaction, or whether a message belongs in one category or another.

The work is usually judged by measurement. How accurate is the forecast? How often are false positives and false negatives happening? Does the model still perform when conditions change? Is the data current enough to support the decision?

Predictive AI is valuable when decisions need a probability, score, ranking, classification, forecast, or warning signal. It should not be treated as certainty. A prediction is a signal for judgment, not a substitute for understanding the operating context.

Best use

Forecasting, scoring, classification, anomaly detection, prioritization, and risk estimation.

Generative AI: what can we create?

Generative AI creates new material. It can draft a paragraph, summarize notes, produce an image, write code, generate test cases, translate a message, propose a plan, or turn rough context into something easier to review.

That makes it powerful, but it also changes the review problem. A generated answer can be fluent and still be wrong. A generated summary can sound reasonable while leaving out an important detail. A generated plan can be useful but still need source checks, constraints, and human judgment before it becomes action.

The strongest use of generative AI is not to replace thinking. It is to reduce the cost of producing a first version, exploring alternatives, and making complex information easier to inspect. The final standard is still whether the output is accurate, appropriate, and tied to the right sources or workflow.

Best use

Drafting, summarizing, rewriting, translating, prototyping, explaining, and generating reviewable options.

Conversational AI: how do people work with it?

Conversational AI is often confused with generative AI because modern chat systems use generative models. But the conversation layer is its own design problem. It is the interface between a person and a system.

A good conversational system does more than produce a pleasant answer. It understands what the user is trying to do, retrieves the right context, asks for missing information when needed, uses tools safely, explains what it found, and makes the next step visible.

This is where many operational systems can become much easier to use. A person should be able to ask a practical question, inspect the evidence, stage a next step, and review what will happen before anything important changes.

Best use

Question answering, guided workflows, command interfaces, support, training, search, and operator assistance.

The categories overlap

In practice, these categories often work together. A conversational system may use retrieval to find evidence, a predictive model to score risk, and a generative model to explain the result in plain language. A planning tool may predict likely demand, generate staffing options, and let a manager refine the plan through conversation.

The overlap is useful as long as the boundaries stay visible. The system should know when it is estimating, when it is generating, and when it is guiding a person through a workflow. Those jobs should not be blurred into one vague claim that "the AI decided."

When the categories are visible, the design can assign the right guardrails. Predictions need measurement. Generated content needs review. Conversation needs grounding, permissions, and clear handoff to action.

Boundary rule

Prediction, generation, and conversation can support each other, but they should not hide each other.

Prediction needs testing

Scores and forecasts should be measured against outcomes, monitored for drift, and treated as probabilistic.

Generation needs review

Drafts and summaries should be checked against sources, constraints, tone, and intended use.

Conversation needs grounding

Dialogue should stay connected to retrievable context, permitted tools, and visible next steps.

A practical way to evaluate an AI system

When I look at an AI system, I try to separate the interface from the intelligence and the intelligence from the authority.

First, what is the system doing? Is it predicting, generating, conversing, retrieving, controlling a workflow, or some combination of those? Second, what evidence or data supports the output? Third, what happens after the output appears? Is it only a suggestion, a staged draft, a scored warning, a routed task, or an action that changes something real?

That last question is where design discipline matters. A system can be helpful without being allowed to act silently. The more meaningful the action, the more the workflow should preserve review, permission, audit, and rollback.

Design question

What kind of AI is this, and what authority does it have after it produces an answer?

How this connects to operational intelligence

This distinction is central to Evidence-Backed AI. Predictive systems can surface signals. Generative systems can turn context into reviewable material. Conversational systems can make the workflow easier for people to use.

The operating standard is still the same: keep evidence visible, keep boundaries clear, and keep authority inside the governed workflow. AI can estimate, draft, explain, recommend, and guide. Important action still needs the right context, permission, and record.

That is the difference between adding AI as a feature and designing AI into a system people can actually trust.

Next pattern

Governed AI workflows: how meaningful actions move through staging, validation, approval, execution, and audit.

About the author

Rodney Herrmann is a systems architect and engineering leader focused on resilient operational systems, emergency communications, automation, data, and governed AI workflows. His work emphasizes practical systems that preserve context, surface evidence, and keep high-impact actions visible and reviewable.